Page 220 - Machine Learning for Subsurface Characterization

190 Machine learning for subsurface characterization

NMR T2 distribution (Fig. 7.1, Track 6) depends on fluid saturations, fluid
mobility, pore size distribution, and surface relaxation, which depends on the
mineralogy of the formation. These physical properties that govern NMR T2
also influence the conventional “easy-to-acquire” logs, namely, neutron,
density, resistivity, dielectric, and GR (Fig. 7.1, Tracks 1–4). These
conventional logs are sensitive to pore size distribution, fluid saturation, and
mineralogy. We use deep neural networks to identify and extract the
complex relationships between these conventional logs and the NMR T2
distribution that cannot be easily identified/quantified by any other
mechanistic models. Deep neural networks cannot find the mechanistic
model but can find a nonlinear higher-order functional mapping that
describes the physical relationships between the conventional logs and NMR
T2 distribution.

3 Data acquisition and preprocessing
3.1 Data used in this chapter
A set of well log data were retrieved from a shale petroleum system comprising
seven formations as shown in Fig. 7.1. The top three formations (F1–F3)
constitute a shale formation, and the bottom four (F4–F7) are dolostone
interbedded with clay-rich conglomeratic dolo-mudstone of Devonian age.
F1 is an upper black shale, F2 is a middle sandy siltstone, and F3 is a lower
black shale. Formations F1 and F3 are hydrocarbon source rocks are
organic-rich with total organic carbon ranging from 12 to 36 wt%. The clay
mineral content is dominated by illite and quartz. Formation F2 is the
hydrocarbon-bearing reservoir and has a low total organic carbon (TOC)
content ranging from 0.1 to 0.3 wt%. Variation in formation mineral
compositions leads to changes in the pore structure, grain texture, and
surface relaxivity. These characteristics along with fluid saturations and their
distribution in the pore network govern the NMR T2 distribution. Mineral
compositions and fluid saturations (Fig. 7.1, Track 5) in the seven
formations were obtained by numerical inversion of resistivity, neutron,
density, gamma ray (GR), and dielectric logs using the mineral inversion
module in TECHLOG. Our goal is to apply the four neural networks to
process 10 inversion-derived logs, including 7 mineral content logs and 3
fluid saturation logs, namely, bound water, free water, and oil, to synthesize
the NMR T2 distributions discretized into 64 T2 bins, which approximate
the fluid-filled pore size distribution of the shale petroleum system. The 10
inversion-derived logs were obtained from conventional logs and will be
referred as the inversion-derived logs. Therefore, the deep neural networks
are trained to relate 10 features (inversion-derived logs) with 64 targets
(NMR T2 distribution discretized into 64 T2 bins).